Claude Code vs Cursor vs OpenAI Codex: Which Coding Agent Costs Less?

Claude Code, Cursor and OpenAI's Codex are the three coding agents most developers are choosing between this year, and all three publish a $20 entry price. That is where the similarity ends. One sells a session allowance, one sells message counts, one sells a dollar pool metered at API rates — and every one of them, once you pass the included amount, charges you the same thing: the model's API price per token.

So "which one costs less" is a question with a real answer, but not the one the pricing pages invite. It is not the sticker. It is what one agent task costs on the model you actually pick, times how many tasks you run in a month, and which published tier that lands in. We priced that with the same maths our AI Agent Cost Calculator runs, at the official token rates checked in September 2026.

Short version: for a daily driver — about twenty six-step tasks a working day — the API-equivalent is $36 to $90 a month depending on the model, so a $20 plan will run dry and a $100 tier will not. The vendor matters less than the model you leave it set to.

CalculatorAI is not affiliated with Anthropic, Cursor (Anysphere) or OpenAI. Prices and plan terms are quoted from the vendors' pricing pages as checked on the dates given below and change often; treat them as a snapshot, not a quote.

The pricing pagesThree ways to sell the same tokens

Each product bundles the agent with a subscription, and each subscription meters usage differently. That difference decides whether "you'll run out by Wednesday" or "you'll get a bill in arrears."

Three details on that table matter more than the prices:

• Claude Code and Codex meter by time window; Cursor meters by dollars. Anthropic sells "at least 5× more usage per 5-hour session than Free" on Pro and 5× or 20× that on Max. OpenAI publishes a table of local messages per five hours by model — roughly 25 to 200 on Plus for GPT-5.6 Terra, 250 to 2,000 for the cheaper Luna, and five times those on the $100 Pro tier. Cursor does not publish an included dollar figure at all; its documentation says only that different models "affect how quickly your included usage is consumed," and offers a rule of thumb — daily agent users "typically $60–$100/mo total usage."
• All three route the overflow to API rates. Anthropic's extra usage is "at standard API rates." Cursor's on-demand usage is "at the same API rates with pay-as-you-go billing." Codex sells additional credits and prices API-key use "based on API pricing," with a rate card of 5 to 250 credits per million input tokens depending on the model. Past the allowance, the three vendors are selling the same commodity.
• The model is the variable you control. Every tool lets you pick, and the choice moves the cost of the same task by 2× to 4× — more than any difference between the vendors' plans.

That is why the rest of this article prices tasks on models, then asks which tier absorbs them.

One taskWhat a coding-agent loop actually costs

A coding agent is not a chat message. It is a loop: read the file, propose a change, run the test, read the output, patch again. Every step resends the whole conversation so far — the system prompt and tool definitions, your request, and everything the agent has read and written since. That growing context is the cost, and it is why an agent bill is dominated by input tokens, not by the code it writes.

We modelled one task as a loop with a 2,000-token system prompt (instructions plus tool definitions, cached), a 300-token request, 350 output tokens per step and 600 tokens of tool result per step — a file read, a diff, a test log. Two shapes: a 6-step edit and a 14-step debug session.

Two things to read off that table. First, the mid-tier models are nearly interchangeable on price: Sonnet 5 and GPT-5.6 Terra are within 7% of each other on both shapes. Second, the 14-step task costs four times the 6-step task, not 2.3× — because each extra step carries everything before it. A debug session that wanders is expensive in a way a long chat is not.

One monthThe API-equivalent of three real workloads

Multiply the task cost by how many tasks you run. We priced three intensities over 22 working days: 5 tasks a day (an occasional helper), 20 a day (a daily driver), and 50 a day (all day, with several agents or automations running).

Switch the workload to 14-step tasks and every cell roughly quadruples: a daily driver on Sonnet 5 goes from $36 to $146 a month, and on Opus 5 from $90 to $364. The shape of your tasks — how many tool calls before the agent is done — is worth more than the vendor's discount.

The fitWhich tier each workload lands in

Put the two tables together. For each workload, the cheapest published tier whose price is at least the API-equivalent is the one that could contain it; anything below it will overflow into extra usage, credits or on-demand billing.

The pattern is plain once it is laid out:

1. 01At the light end, all three $20 plans hold. Five tasks a day on a mid-tier model is under $10 of tokens. The choice between them is about the editor and the workflow, not the bill.
2. 02A daily driver on a mid-tier model sits in a gap. $36–$39 is too much for any $20 plan and far less than $100. Cursor is the only vendor with a tier in between — Pro+ at $60 — which is exactly the segment its own documentation describes ("daily agent users typically $60–$100/mo").
3. 03Opus 5 or Sol as the default model is what pushes people up a tier. Twenty tasks a day on Opus is the same $90 as fifty a day on Sonnet. The premium model costs roughly two to two and a half times the mid-tier on identical work.
4. 04Above $200 a month the vendors converge. Every product's top tier overflows into API-rate billing, so the bill is the API bill plus whatever the subscription included. At that point the honest comparison is the API itself — see Claude Pro vs Max vs API and ChatGPT Plus vs Pro vs API for the break-even arithmetic on each side.

The catchesFour line items the sticker hides

The modelSonnet-class or Opus-class is the real decision

Every one of these tools lets you set a default model and switch per task. The cheapest configuration on all three is the same: a mid-tier default, the premium model on demand. A team that leaves Opus or Sol as the default is choosing to pay roughly two to two and a half times as much for the majority of tasks that did not need it — that is the model-routing decision that dominates every AI bill, and coding agents are where it shows up first.

The decisionA rule that survives the next price change

If your usage is genuinely unknown, run a month on the $20 tier of whichever editor you prefer, then read the usage page and put the token counts into the AI Subscription vs API Calculator. It answers the question these pricing pages avoid: at your volume and your model, is the subscription a discount or a convenience fee?

Frequently asked questions

Which is cheapest: Claude Code, Cursor or Codex? For light use, all three start at $20 a month and none of them runs out. For a daily driver on a mid-tier model, the work is worth about $36–$39 of tokens a month; Cursor's $60 Pro+ is the only tier priced between $20 and $100. For heavy use on premium models, every product overflows into API-rate billing, so the bill converges on the API bill.

What does one coding-agent task cost? At September 2026 API rates, a six-step task with a cached system prompt costs about $0.06 on Claude Sonnet 5 or GPT-5.6 Terra, $0.12 on GPT-5.6 Sol and $0.15 on Claude Opus 5. A 14-step debug session costs about four times that, because every step resends the whole context.

Does Claude Code's $20 Pro plan include enough for daily coding? Anthropic sells usage per five-hour session rather than a token budget, so it depends on how bursty the day is. As a ceiling: twenty six-step tasks a day on Sonnet 5 is about $36 of API usage a month, which exceeds the $20 plan price, so a daily driver should expect to hit limits or enable extra usage at API rates.

How does Cursor bill when the included usage is gone? Cursor's documentation says you can add on-demand usage and "continue at the same API rates with pay-as-you-go billing," billed monthly in arrears, or upgrade the plan. The included amount per plan is not published as a dollar figure.

Is the Codex API cheaper than a ChatGPT plan? API use is billed per token at the model's rate, with a credits table from 5 to 250 credits per million input tokens. Whether it beats Plus or Pro depends on volume and model — our ChatGPT Plus vs Pro vs API comparison works through the break-even.

Why do agent tasks cost so much more than chat? Because an agent resends the whole conversation on every tool call. A 14-step task sends about 120,000 tokens of input for under 5,000 tokens of output; a chat question sends a few hundred. Input volume, not code volume, sets the bill.

Sources and calculation method

• Anthropic, Claude pricing page (Pro $20 monthly / $17 annual; Max from $100 with 5× or 20× Pro usage; extra usage "at standard API rates"), checked September 16, 2026.
• Cursor, "Pricing" documentation (Pro $20, Pro Plus $60, Ultra $200; two usage pools; on-demand usage "at the same API rates"; Teams $40 Standard / $120 Premium; $0.25 per million tokens Cursor Token Rate on third-party models for Teams and Enterprise), checked September 16, 2026.
• OpenAI, Codex pricing documentation (included in Free, Go $8, Plus $20, Pro from $100 with 5× and 20× tiers, Business $20–$25 per user; local messages per five hours by model; additional credits; API use at API pricing with a 5–250 credits-per-million-input-tokens rate card), checked September 16, 2026.
• Model rates: official standard API prices as recorded in CalculatorAI's model table, verified September 9, 2026 — Claude Sonnet 5 $2/$10, Claude Opus 5 $5/$25, GPT-5.6 Terra $2/$12, GPT-5.6 Sol $4/$20 per million input/output tokens, cached input at one tenth of the input rate.
• Method: every task figure is produced by the same function the AI Agent Cost Calculator runs, with the parameters stated in the article; the script is kept as an editorial artifact. Assumptions bias the numbers downward: no retries, no images or long files in context, a cached system prompt, and 22 working days. Real sessions with larger repositories or uncached tools cost more.